Raw robot data in.Training-ready datasets out.
Free 100-episode pilot · Results in 48 hours
Annoroid turns teleop logs, rollouts and egocentric video into data you can train on.
Task, in the dataset's own words
Place the coffee capsule inside the capsule container, then place the cup onto the center of the cup tray, then push the Hot Water and Travel Mug buttons.
- Pick up the capsule
- Open the capsule container
- Place the capsule inside
- Close the container
- Place the cup on the tray
- Push Hot Water and Travel Mug
- Return to rest
Footage: ALOHA static coffee (LeRobot / ALOHA project, MIT). Step labels by Annoroid.
In
- LeRobot
- HDF5
- ROS 2 bags / MCAP
- Egocentric video
Software does the first pass. Trained human reviewers verify every episode.
Out, per episode
- Cleaned episodes
- Quality score per episode
- Rejection reasons
- Sub-task segmentation
- Language task descriptions
- Failure labels
Works with the formats and storage you already use
- LeRobot
- HDF5
- ROS 2 bags
- MCAP
- Egocentric video
- MP4 + IMU
- Parquet
- AWS S3
- Google Cloud Storage
- Cloudflare R2
- Hugging Face Hub
Bringing order to the chaos of
Raw robot data
Every collection run produces episodes you can't train on as-is. Teams lose weeks to one-off scripts and scrubbing video before a single training run.
ep_0007 idle start .......... 3.4 s
ep_0007 wrist_cam ............ frozen
ep_0012 operator_error ....... regrasp
ep_0019 ok
Noisy.
Failed grasps, idle starts, operator corrections and frozen sensors sit next to the good demonstrations.
Three steps to a dataset
from the raw logs you already have.
- 01
Connect your data
Point us at a bucket or upload through the API. LeRobot, HDF5, ROS 2 bags / MCAP or egocentric video. We read your format; you don't convert anything.
- 02
We process and verify
Automated checks flag sync errors, gaps, frozen sensors and idle time, then propose segments and labels. Trained reviewers check every episode and correct what the software got wrong.
- 03
Receive a training-ready dataset
Same format you sent, plus a quality report and labels, back in your storage or ready to download. Poll the job or get a webhook when it's done.
Real outputs.
Every episode comes back with all six.
Delivered as files next to your data (JSON / Parquet) and through the API. Your originals are never overwritten.
- 01
Cleaned episodes
Idle time trimmed, boundaries fixed, streams aligned. Your originals are never overwritten.
- 02
Quality score per episode
One score plus the individual checks behind it, so you set your own threshold.
- 03
Rejection reasons
Episodes we'd drop, each with a reason code such as sync_drift, operator_error or occluded_camera.
- 04
Sub-task segmentation
Start and end frame for each step of the task.
- 05
Language task descriptions
A task instruction per episode and a short instruction per segment.
- 06
Failure labels
Success or failure, plus where it failed and the failure type.
Fixed turnaround
48 hours for pilots. Production turnaround is agreed per order, in writing.
Acceptance criteria up front
Acceptance criteria agreed with you before each pilot, and free re-work if delivered data doesn't meet them.
Free re-work
Miss the turnaround or the acceptance criteria and we redo the affected episodes at no charge.
Real automated-check output for 5 episodes of lerobot/aloha_static_coffee, as JSON and Parquet, plus the quality report for all 50.
Our checks, run on
Public robot data
Fleet data you can't search is data you can't train on.
Annoroid Index
Every episode your robots record, indexed by what happened: task, steps, outcome, failure type and sensor health. Watch it run on 2,131 real episodes from public datasets, then ask your own questions.
Clean episodes from G1 humanoids
checks.all = passed
robot ~ "g1"
→ 777 of 2,131 real episodesSend us 100 episodes.
Judge the output yourself. The pilot is free, results come back in 48 hours, and there's no contract.